ELIZA and the Illusion of Memory (1966)
Joseph Weizenbaum's ELIZA, created at MIT in 1966, used simple pattern matching and scripted responses to simulate a psychotherapist. ELIZA had no memory — each response was generated purely from the current input with no reference to previous turns. Yet users frequently reported feeling that ELIZA truly understood and remembered them, a phenomenon Weizenbaum called the "ELIZA effect." The illusion of memory proved powerful even before memory existed.
Early Chatbot Systems and Session State
Through the 1970s and 1980s, conversational systems began maintaining session state — a temporary record of what had been said within a single conversation. PARRY (1972), which simulated a paranoid schizophrenic, tracked assertions made earlier in the conversation to maintain consistency. These systems maintained within-session coherence but had no cross-session persistence.
The Rise of Rule-Based Dialogue
1990s commercial chatbots, deployed in customer service applications, introduced more sophisticated dialogue state machines. Companies like SmarterChild (AIM, 2001) allowed users to set preferences that persisted across sessions — rudimentary but groundbreaking personalization. This era established the principle that conversation history has ongoing commercial value.
Neural Networks and Context Windows
The deep learning era introduced the concept of the context window as the mechanism for conversation memory. Early transformer models processed the full conversation history within a single forward pass, limited to a few thousand tokens. The context window became the fundamental constraint: everything inside it was "memory," everything outside was gone. GPT-3's launch in 2020 brought this architecture to mainstream awareness with a 2,048-token context window.
Modern LLMs and the Memory Problem
GPT-4 and Claude expanded context windows to 128k and 200k tokens respectively, but the fundamental reset problem remained. Every new conversation starts from zero. The AI has no memory of what you discussed yesterday, last week, or last year — unless you explicitly re-provide that context. This limitation is not an oversight; it's an architectural consequence of stateless inference. Solving it at scale is the defining engineering challenge of the AI memory category — and the market opportunity that ChatHistory.com represents.